Deep learning technology detects snow cover on PV systems and calculates energy loss – SPE

September 2, 2024 Lior Kahana

Researchers from Canada’s University of Waterloo have developed a new automated model for real-time detection of snow cover on PV panels and estimation of energy loss. The model is based on image processing and deep learning, using a convolutional neural network (CNN) and a time-lapse sequence of images to identify the snow-covered area compared to the clean area.

“In cold climates, where snow cover can last for days or weeks and cause up to 34% energy losses, timely detection is crucial for maximizing the energy yield of PV systems to minimize these losses,” the researchers said. “The ability to autonomously detect snow on PV systems is not only crucial for maximizing energy yield, but also for maintaining the integrity of solar panels and ensuring a stable power supply.”

The proposed method consists of five steps. First, a dataset of 250 images of PV panels against different backgrounds is obtained. The images are then processed using various image transformation techniques to expand the dataset to 800 images. In this step, the research group manually created a binary black-and-white mask for each image, with white pixels representing the PV panels and black pixels the background.

At the next level, all pairs of images and masks are fed into U-Net CNN, a convolutional neural network specifically designed for image segmentation tasks. The model is asked to view the dataset and identify PV panels. It then compares the results with the ready-made masks, and in case of errors, the images are propagated back through the system.

In the fourth step, the model is validated against the ground truth. In the final step, the energy loss due to snow accumulation is estimated based on the differences in the opacity of the white pixels in the current image and the designated image without snow. To calculate the energy loss, a mathematical model from previous literature is used.

“The presence of snow in the background can lead to confusion as to whether it is on the panel or not, adding an additional layer of complexity to the accurate estimation of snow cover,” the team explains. “To address these challenges, a reference image of the PV scene before the snowfall was used. This reference image must be from the same distance and angle as the test image and must be captured when there was no snow on the PV panels. Uniform camera settings for both reference and test images were essential to minimize camera-induced differences.”

The new method was tested on six new images on which the model had not been trained. The system’s snow identification and energy loss were compared with the capabilities of popular image processing-based segmentation methods, such as Canny edge detection, K-Means-based segmentation, and image thresholding.

The results indicated that the new method achieved an average error of 2.88% and a Dice coefficient score of 0.91. The dice score is calculated as the ratio between the predicted and ground truth regions. In compression, K-means clustering achieved an average Dice score of 0.47, Canny edge detection got 0.60 and the threshold method showed a score of 0.63.

“Energy losses were also estimated from the images with average energy losses of 1.5 kWh/m2/month from six different PV arrays,” the scientists explained. “The model could reliably predict energy losses due to snow accumulation with an average error of 0.05 kWh/m2/month. The maximum energy loss was 0.23 kWh from a large PV array system with an area of ​​117 m2. Analysis of the impact of the duration of snow cover on energy loss showed a savings potential of 0.13 kWh/m2 if the snow cover was cleared in a timely manner.”

The method was presented in “Leveraging deep learning for real-time snow cover detection and energy loss estimation for solar panels”, published in Applied energy.